Extreme parkour with legged robots
extreme-parkour.github.io
extreme-parkour.github.io
So that's what it takes. That's so much simpler than the way Boston Dynamics does it, working out all the dynamics in simulation first. It's amazing to see this done from vision to actuators in one net. It's now much less of a mystery how animals managed to evolve the ability to do that.
(I had a go at this problem years ago. I was trying to work out the dynamics from first principles, knowing it would be somewhat off. Then use some self-tuning (a predecessor to machine learning) to fine-tune the thing. Got as far as running up and downhill in 2D in the early 1990s. About one hour of compute for one second of motion.)
A very recent development (which this work builds on) is the idea of "online adaptation". It essentially involves doing the training in two stages:
1. You add a variety of dynamically-varying environmental effects to your simulator, by randomly altering parameters such as ground friction, payload weight distribution, motor effectiveness, and so on. You give the motion controller perfect knowledge of these parameters at all times, and let it learn how to move in response to them.
2. Then, you remove the oracle that tells the controller about the current environmental parameters, and replace it with another neural network that is trained to estimate (a latent representation of) those parameters, based on a very short window of data about the robot's own motor commands and the actual motion that resulted from it.
All of this can be done in simulation, many times faster than real-time. But when you transfer the system to a real robot, it adapts to its environment using the estimated parameters, without any of the networks needing to be re-trained. This ends up making it pretty robust to difficult terrain and perturbations. It also has the benefit of papering over subtle differences that arise between the simulated and real-world dynamics.
This paper adds a lot of additional refinements to the same basic idea. In the first stage, the system is given perfect knowledge of its surrounding terrain and the locations of some preselected waypoints, and learns to follow them. The second stage replaces those inputs with estimates derived from an RGB+depth camera.
Where do you need to be ?
Where is debatable, but you definitely need to be there by end of Q3 or we miss P&E targets.
> So that's what it takes. That's so much simpler than the way Boston Dynamics does it, working out all the dynamics in simulation first.
I mean, it sounds like they have had to work out all the dynamics, and simulate them.
Admittedly they don't have to do it on-robot in real time any more.
"Parkour is a popular athletic sport that involves humans traversing obstacles in a highly dynamic manner like running on walls and ramps, long coordinated jumps, and high jumps across obstacles. This involves remarkable eye-muscle coordination since missing a step can be fatal. Further, because of the large torques exerted, human muscles tend to operate at the limits of their ability and limbs must be positioned in such a way as to maximize mechanical advantage. Hence, margins for error are razor thin, and to execute a successful maneuver, the athlete needs to make all the right moves. Understandably, this is a much more challenging task than walking or running and requires years of practice to master. Replicating this ability in robotics poses a massive software as well as hardware challenge as the robot would need to operate at the limits of hardware for extreme parkour."
Here's an example of real parkour: https://www.youtube.com/watch?v=5lp1oS0vXg0
They aren't doing parkour, let alone "extreme" parkour.
I'm not trying to diminish the accomplishment with respect to the state of the art of robotics, just bring some reality to the comparison against what humans are capable of.
More impressive would be adaptation to obstacles without clearly delineated edges, sticking landings on uneven/moving landing sites, and especially avoidance of landing sites which appear incapable of supporting its weight properly, particularly if it could do it well enough to generalise to novel courses.
The video hints the model may be able to do this to some extent (the high jump does show some apparently necessary compensating movement to avoid slipping off), but doesn't really demonstrate it.
E.g. A sleeping dog that is motionless, a valuable item, wet concrete.
I suppose I'm going out-of-scope, though.
The one pictured appears to be $14k
But with the parkour on top of city buildings that is more extreme than what they portray on rooftop chases in a Bourne Identity movie... hmm... I think that is another level. There is a good chance you will die if you do that alot over 10 years. It probably requires some letting go of ego to the point where you are happy to die early. Like the free and solo climbers.
[1] Forgive me if you are a great skiier this is my dumb interpretation!
With Skiing, you have to make yourself quite vulnerable. It’s instinctual for us to want to close ourselves up to protect our organs when scared. But you can’t ski well or difficult terrain without opening your body up, relax your arms and expose your vulnerability. This is especially scary at first when tree skiing. When you do relax it’s amazing.
I think this is why some like snowboarding more. You’re facing sideways and you feel less exposed. You can get away with being closed up a little more and while it’s bad form, you never really have to snap out of it the way way you do on Skis.
The feeling over fear when you want to ask someone out, then you do it and it works out, that is the spice of experience. Like wise when it doesn’t work out, it can suck but it’s also part of the spice.
They are essentially one “organism”. They are effectively the same height, same weight, same battery life, same way of processing information, same capabilities.
There is zero uniqueness. Like a single machine with a thousand eyes and arms.
If they develop uniqueness it would get wild, like the movie rogue one…
If a robot is damaged it would be interesting though. Would it adapt ? If it adapts would it be part of the same fleet? Could it receive the same “updates”, or would it be an individual? It would the able to do certain things others bots of the same model could do…
I feel the way the robots do it is similar to humans. Considering that's all emergency behaviours (not trained with animals' or humans' motion data) I believe that robots will exceed humans by far in few years.
Like the whole point of saying “extreme” parkour is to boost engagement from pedantic analytic people like us talking about the hyperbolic title choice
I hate what the attention economy has become. Old man yells at emoji...
In terms of landing on the back: if I remember correctly, one of these prototypes--from Boston Dynamics?--is able to rotate its legs 180 degrees to push off the ground when landing on the back.
For the other, legs are more versatile than wheels, especially when it comes to uneven terrain.
https://spectrum.ieee.org/wheels-are-better-than-feet-for-le...
Outragously expensive is not correct if you ask me.
What would be really interesting if some way of motion turned out to be much better than what we find in nature or manufactured for non-autonomous machines.
Or is it some nuance of metals vs carbon-based joints?
(Edit: although come to think of it, in the animal world, jump height is much more constrained by 9,81 m/s2 than by body size)
What really matters is the energy density of your storage medium, which is the same to an order of magnitude for any system based on chemical energy. So both a flea and an elephant (along with battery-powered robots) can jump about a meter or two in height.
https://www.youtube.com/watch?v=mvHXwTa5-DA
(But for all but extreme designs I fully agree with you)
Delivering explosive payloads is a solved problem. Detecting humans is a solved problem. The hard part is reliably distinguishing enemy solders from friendly soldiers or civilians.
These could swarm in the back door.
This is why, despite being technically feasible, we haven't seen drone swarms yet. You can easily build 1000 drones, but you can't have 1000 channels of live video over radio back to 1000 drone operators.
It takes about 15 rounds of 155mm artillery per battlefield casualty. Each shot costs about $5k.
Robots like this might cost the same as a single artillery shot but have the ability to get multiple kills each.
It’s a cold arithmetic, but I guarantee you someone at DARPA is thinking and planning the future of warfare in these terms.
But some googling indicates this is about right, and prices have even gone up since the Russian invasion.
A typical round weighs around 40 kilos or 100lbs. For reference, from a quick web search, as commodities:
- 100lbs of TNT would be around $500
- 100lbs of stainless steel is around $250
I'm not sure what materials go in (I'd assume better explosives and worse steel), but I think $250-$500 is a good estimate for price of raw commodities (not including anything beyond that).
I would at least double that for having those shaped into an artillery shell, and add the cost of non-commodities (like a detonator), and you're easily at $1000. I would at least double that for the raw cost of distribution, sales, shipping, and logistics, and other overhead. We're probably at $2000.
At that point, toss in 30% profit margin for everyone along the way (30% to manufacturer, 30% to distributor, etc.). You're probably now around $3500.
I think that's a bare minimum baseline for what it would take to get a very, very dumb shell. This goes up if you want:
- Anything at all fancy or high-tech
- War profits
- NREs covered (building a factory, which might be idle 95% of the time and spin up for wartime)
- Military inefficiencies
- Regulatory compliance, quality inspection, etc.
Etc.
You can hit $5k very easily.
Here's a military news article covering the same umbrella grant program this research was funded under: https://www.c4isrnet.com/2022/06/23/darpa-adding-common-sens...
That's offense not defense. It worsens the national security of your opponent but doesn't strengthen your own.
You don't defend territory with static defense's or force fields or walls.
Drones and robots made attacks easier, did that help national security?
Countries like Russia or China won't attack because of nuclear weapons but groups like ISIS don't care.
Large flying drones make patrolling a border easier. You don't have to worry about the pilot's fatigue level. Only fuel/battery.
The war in Ukraine clearly shows that smaller drones also help in attacking and defending. Attacking, by dropping bombs or spotting targets. And defending, via spotting incoming enemy forces; spotting enemy artillery that is shooting at you, etc
Yes because the enemy will have them eventually regardless of whether you develop them. Call it a Prisoner's Dilemma if you will, but it's the reality. Besides, unlike nuclear weapons, warfare with autonomous systems is not particularly more cruel than WW1 or WW2 style warfare.
But the word "defense" is much wider than that. Is the US supplying offensive weapons to Ukraine contributing to the national defense? Even if you feel the US should not be doing that, I hoe you can see the logic of people who do frame it that way.
And Ukraine's in a defensive war against an invader, and for that they have to go on the offense.
(Of course the US renaming the Department of War to Department of Defense in 1947 was 100% propaganda, or to be more charitable, aspirational. There is no question that it has been used offensively).
In a war the line between offense and defense is blurred but the national security is at level zero.
It's not. It's contributing to European security, and as such is necessary for the US to reaffirm their suzerainty over Europe, but it's not really about US defense and even less about National Security. (I'm glad they do btw, because we European powers would have left Ukraine fall after deciding it would be too expensive to help them…)
If you consider “national security” to be a meaningless buzzword and not an actual concept, you can say that. It's as accurate as saying “US sports results at the Olympics is a national security issue via soft power”.
Pretty much everything is “national security” by that standards.
> Being a global superpower has many benefits for US citizens.
Sure, but most of them don't have anything to do with national security.
Consider terrain such as the mountain ranges along the side of California. Or the entirety of Japan.
Wheeled and tracked vehicles generally have problems in that terrain.
Legged robots don't. Suddenly your troops intended to operate in mountainous terrain have a reliable robot donkey to use as a packmule.
Many countries (US, Germany, France, Italy, etc) still use horses and donkeys for transport in bad terrain.
This is the tragedy of the commons. I'd rather see nobody making these advances, but if someone has to it'd better be people on my side. Can't stop progress :(
Yeah, screw GPS, trauma surgery improvements and better prosthetics, IFF systems, radar/sonar, Internet, ToR, Epipens, and duct tape. /s
Really, I get having a generally anti-war position and wanting to spend money on other shit but pretending like there are no positive externalities to defense spending is childish.
There’s a lot of wholes in the plot, but the overall premise and scenario seem eerily plausible
You said yourself that other nations can and will make this too. Avoiding a decrease of national security is the same as improving national security.
Is this not the case with all the major R1 institutions?
Walt Rostow may no longer be deciding on troop levels from his office in the MIT Economics department (that cause MIT to "divest" its overtly military work into figleaf organizations) but still, after that, ARPA (later DARPA) poured hundreds of millions into the institution.
They basically paid for my education there and I am not a US citizen and never worked on anything military-related, either while in school or since.
It's certainly true that many, many universities are deeply embedded with the DoD, defense contractors, and other weapons manufacturer ecosystems. Certainly CMU isn't exceptional here. But I think it's very important to keep in mind the reality of these programs when they come up in the news, like this article, which is why I left the link to how "the other side" sees these types of research programs.
Fun fact, in the mid-80's early 90's the "industry" was "getting ahead" of the Department of Defense in terms of capabilities that could be deployed in an adversarial way. One of the most visible outcomes was the first crypto wars. (source code became ITAR controlled, emailing a perl script that implemented the RSA algorithm to someone outside the country was an ITAR violation, and we joking suggested illegal immigrants get that code tattooed on their body so that it would be illegal to deport them without a license from the Dept. Of Commerce)
Getting an inertial navigation unit w/sensors that weighed less than 10kg (22 lbs) was code word level secret stuff, because beyond the line of sight missiles are a thing.
Anyway. the US DoD (like the defense departments everywhere) realized they needed to be more engaged with R&D labs if only to see what progress they were making so that they could anticipate whether or not that progress might show up as a threat in some way. On the plus side that freed up some money that would have been in black budgets for universities that were doing similar research anyway and as a funder, the source, in this case the DoD, generally gets a non-exclusive perpetual right to use any resulting IP out of that sort of funding arrangement, even when patented.
Bottom line, researchers are gonna research, so make sure what the future holds is not gonna show up unexpectedly on the battlefield or in other covert ways.
All that said, as someone who has been involved in building this "scale" of robot (under 1m, self powered) the progress has been freakin' unbelievable. In part because of the fact that you can put a supercomputer on one that weighs less than 500 g and runs for 8 hours on a battery that weighs less than 1 kg. Sensors, video, real time image processing. All pretty stunning.
The Ukranian use of off the shelf drones, and their rapid development of weaponized "drone munitions" tells me that the DoD is correct in wanting to keep track of what is going on with robots that can do "extreme parkour."
I was on dog walk duty and always held the lead close, but I knew they were worried I wasn't so conscientious and would let my dog friend run amuk on their garments.
So I was trying to imagine an example of a creature that would horrify me similarly. Like perhaps if people had giant BEETLES as pets, and thought it was all grand, and I was like "what the FUCK are you on, my friend?", and I was pretty happy with that.
But I think a closer cut to how disconnected this is from what seems appropriate, and normal, and reassuring, would be these robots. Clearly, some people don't really understand how horrifying this is, particularly with elevator music as the background. And if you want to know how upsetting the elevator music is as a background to set tone, I strongly urge you NOT NOT NOT to google "crab club" on youtube.
"The Hound half rose in its kennel and looked at him with green-blue neon light flickering in its suddenly activated eyebulbs. It growled again." "THEY sent A SLAMHOUND on Turner's trail in New Delhi, slotted it to his pheromones and the color of his hair. It caught up with him on a street called Chandni Chauk and came scrambling for his rented BMW through a forest of bare brown legs and pedicab tires. Its core was a kilogram of recrystallized hexogene and flaked TNT. He didn't see it coming. The last he saw of India was the pink stucco facade of a place called the Khush-Oil Hotel." "Ng Security Industries Semi-Autonomous Guard Unit #A-367 lives in a pleasant black-and-white Metaverse where porterhouse steaks grow on trees, dangling at head level from low branches, and blood-drenched Frisbees fly through the crisp, cool air for no reason at all, until you catch them."Practically, because legs are superior to wheels when it comes to pursue a human target. It may be superior to flying as well indoor, to open doors, etc. And then there is the energy expenditure and additional mass constraint when airborne.
Research on legged robots is not just for 'fun', it's because they are better at going through obstacles and rough terrain than wheeled vehicles.
Flying can get you almost anywhere, but not quite. It is both very loud and has issues in enclosed spaces.
Wheels/tracks are very efficient at moving mass, but have issues crossing some materials such as exploded building internals like you would find in a war zone.
Walking/climbing is the last bit to get you inside enclosed 'human' spaces such as buildings and tunnels.
The war occurring now in Gaza would be an example area of what future battle would look like.... A target building it hit by a bomb and mostly destroyed. Smaller automated drones quickly fly over the area and assess visible targets and threats for follow up artillery. Multiple mobilized traced/wheeled heavily armored drones would address any residual ground level threats. Upon reaching the target building the larger drone vehicles would release multiple smaller 'walking/crawling' style drones to ascend the rubble, either addressing any remaining targets themselves, or reporting back to a command station for some other group to make a decision on what to do next. The final goal of these walking units is to probe deep into the rubble to ascertain if any tunnels survived the initial bombing, and if so to begin mapping any threats in said tunnels.
Personally,I think we'll get really good at killing robots and they'll get better at killing people (military).
If you watch it then those research legged robots will look rather more sinister... or maybe it's just me. That episode really stuck with me.
These robots won't have any problems pulling the trigger at anyone or anything. They won't hesitate or refuse an order the way Stanislav Petrov did. They won't be trained to have empathy or respect for life. Humanity is the final check against the worst atrocities war demands from the people we send out to kill each other, and devices like these solve that "problem" by dehumanizing the soldier.
Come on man, this is delusional. I want to believe it too but be realistic.
Also the comments on that other story are just sad. I understand war is in the news but I hope it's not all we can think about! Also I'm surprised by all the people who think this is not impressive because "a 10 year old could do this" or whatever, that's actually ridiculous.
I guess HN just isn't the right community to appreciate real hard technology like this anymore. This makes me sad, especially because I don't know of anywhere else that's better.
You get a chuckle before you die